基于多尺度特征融合神经网络的织物表面缺陷检测方法  

Fabric surface defect detection method based on multi-scale feature fusion neural network

作　　者：孙迪 王旭彬 SUN Di;WANG Xubin(School of Film&Television Art,Liaoning Normal University,Dalian 116081,China;School of Computer Science and Information Engineering,Shanghai Institute of Technology,Shanghai 201418,China)

机构地区：[1]辽宁师范大学影视艺术学院,辽宁大连116081 [2]上海应用技术大学计算机科学与信息工程学院,上海201418

出　　处：《辽宁师范大学学报（自然科学版）》2024年第3期331-341,共11页Journal of Liaoning Normal University:Natural Science Edition

摘　　要：织物表面缺陷检测是织物生产过程中必不可少的环节,近年来,使用机器视觉的方法逐渐被提出,但使用该方法存在诸多局限性,致其精度和对未知缺陷的检出率不高.提出了一种基于多尺度特征融合神经网络的织物表面缺陷检测方法,其基于编解码器架构,编码器的特征提取模块使用多尺度卷积核和空洞卷积以感知不同尺寸的缺陷,结合残差并加入了特征增强筛选模块以优化多尺度信息,增强不同特征之间的关联.最终设计了一种融合带权交叉熵和Dice Loss的损失函数以加速模型收敛.实验表明本方法可以有效分割出织物表面的缺陷部分,在AITEX织物数据集上达到了98.96%的准确率以及96.95%的召回率.Fabric surface defect detection is a crucial step in the fabric production process.In recent years,machine vision methods have been increasingly proposed for fabric defect detection.However,these methods have several limitations that affect their accuracy and the detection rate of unknown defects.This paper proposed a fabric surface defect detection method that utilizes a multi-scale feature fusion neural network,leveraging the encoder-decoderarchitecture.The feature extraction module of the encoder utilizes various convolution kernels and dilated convolutions to detect defects of varying sizes.Additionally,it combines residuals and incorporates a feature enhancement and filtering module to optimize the utilization of multi-scale information,enhancing the association between different features.Finally,a loss function is designed that combines weighted cross-entropy and Dice Loss to facilitate model convergence.Experiments show that this method can effectively segment out the defect parts of the fabric surface,achieving an accuracy rate of 98.96%and a recall rate of 96.95%on the AITEX dataset.

关 键 词：织物缺陷检测 深度学习 计算机视觉 神经网络 异常检测 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]